Interpretable Machine Learning Model for Early Mortality Prediction in Septic Patients Using Routine Post-Diagnosis...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Prediction Models

Background:

• Early identification of high-risk sepsis patients is critical for improving survival rates.
• Routine clinical data collected at diagnosis can be leveraged for predictive modeling.
• Developing accurate predictive tools is essential for timely clinical decision-making.

Purpose of the Study:

• To develop and validate a machine learning model for predicting 7-day mortality in sepsis patients.
• To assess the model's performance against traditional scoring systems.
• To identify key clinical predictors of early sepsis mortality.

Main Methods:

• Utilized data from 8729 sepsis patients across four Chinese tertiary hospitals.
• Trained and evaluated seven machine learning algorithms, including Artificial Neural Network (ANN).
• Employed Area Under the Receiver Operating Curve (AUROC), calibration curves, Decision Curve Analysis (DCA), and SHapley Additive exPlanations (SHAP) for evaluation and interpretation.

Main Results:

• The ANN model achieved a superior AUROC of 0.767, outperforming APACHE II (0.710) and SOFA (0.718) for 7-day mortality prediction.
• Key predictors identified by SHAP analysis included Glasgow Coma Scale (GCS), blood chloride, and albumin levels.
• Model performance was validated and demonstrated consistency across training and testing datasets.

Conclusions:

• A robust machine learning model was developed for predicting early sepsis mortality using readily available clinical data.
• The SHAP-based interpretation enhances model transparency, aiding clinicians in identifying at-risk patients.
• This validated tool holds potential for clinical application, facilitating early interventions and improved patient outcomes.